Research map / Label-efficient and robust learning / Recognition and long-tail
Bias In, Bias Out? Finding Unbiased Subnetworks in Vanilla Models
Ivan Luiz De Moura Matos, Abdel Djalil Sad Saoud, Ekaterina Iakovleva, Vito Paolo Pastore, Enzo Tartaglione
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add reweighting of the cross-entropy loss (ablation) | BISE mask objective · BiasedMNIST (rho=0.99) | accuracy: 91.9 -> 96.0 | Improved |
| add the mutual-information term between predicted and true bias (ablation) | BISE mask objective (with reweighting) · BiasedMNIST (rho=0.99) | sparsity: 18.5 -> 20.9 | Improved |
| add the mutual-information term (without reweighting) (ablation) | BISE mask objective · BiasedMNIST (rho=0.99) | accuracy: 91.9 -> 91.1 | No clear effect |
| add the mutual-information term (ablation) | BISE mask objective · BiasedMNIST (rho=0.99) | sparsity: 5.8 -> 18.7 | Improved |
| add finetuning of the extracted subnetwork | BISE · BiasedMNIST (rho=0.997) | accuracy: 90.8 -> 95.9 | Improved |
| add finetuning of the extracted subnetwork | BISE · BiasedMNIST (rho=0.999) | accuracy: 15.7 -> 60.5 | Improved |
| add finetuning of the extracted subnetwork | BISE (last) · Corrupted-CIFAR10 (rho=0.995) | accuracy: 21.94 -> 34.18 | Improved |
| add finetuning of the extracted subnetwork | BISE · CelebA | accuracy: 89.7 -> 91.8 | Improved |
| add finetuning of the extracted subnetwork | BISE · CivilComments | worst-group accuracy: 80.4 -> 81.0 | Improved |
| extract subnetwork with BISE from vanilla model | Vanilla model · BiasedMNIST (rho=0.997) | accuracy: 66.1 -> 90.8 | Improved |
| extract subnetwork with BISE (last) from vanilla model | Vanilla model · Corrupted-CIFAR10 (rho=0.95) | accuracy: 47.18 -> 55.38 | Improved |
| use pseudo bias labels instead of true bias labels | BISE · CelebA | accuracy: 89.7 -> 87.5 | No clear effect |
Closest papers by meaning
- Training Debiased Subnetworks With Contrastive Weight Pruning (CVPR 2023 · similarity 0.92)
- DeNetDM: Debiasing by Network Depth Modulation (NeurIPS 2024 · similarity 0.89)
- Initial Guessing Bias: How Untrained Networks Favor Some Classes (ICML 2024 · similarity 0.86)
- Diffusing DeBias: Synthetic Bias Amplification for Model Debiasing (NeurIPS 2025 · similarity 0.86)
- Enhancing Intrinsic Features for Debiasing via Investigating Class-Discerning Common Attributes in Bias-Contrastive Pair (CVPR 2024 · similarity 0.86)
- NeuronTune: Towards Self-Guided Spurious Bias Mitigation (ICML 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
